fix: close real-model planning feedback and reviewer order gaps

[shenjiecode]

Summary

Closes engineering gaps found by the frozen Qwen real-model comparison; does not claim the quality evaluation has passed.

• Supply the complete five-field scenario template, stable-ID/status rules, and actionable bounded patch/schema errors without exposing raw Git stderr or rewriting patches.
• Validate patch + resulting scenario states + execution selection at planning completion. Feed errors back at most twice within the same isolated Main Session; keep existing fail-closed checks and clean the temporary worktree.
• Enforce Reviewer plan/existing-patch/original-image order before draft access or review submission. Failed image reads remain blocking; clarify that an execution list does not waive required capability coverage.
• Test request recovery separately from Session completion; keep artifact and human quality checks distinct in evaluation tooling.

Verification

Using the lockfile-matching Docker quality image with current release inputs mounted:

• format / lint / typecheck: passed
• unit + integration: 162/162, 26 files
• build + headless E2E: passed
• local acceptance including production Pi paths: passed (live/release blocked)
• tests cover parser-valid role template, rejected premature reads without exposing content, existing patch requirement, failed/zero-image paths, bounded same-Session repair, explicit missing-newline feedback, and recovered request classification.

No credentials, production data or raw evaluation outputs are committed. No target website operations, target remote publication, deployment or release.

Real-model status

The previous 72 attempts and all failures remain intact. New candidate a1bdd68 is frozen for a fresh 72-Session comparison against the original 09dbed0 baseline, using unchanged website-derived fixtures/rubric and qwen3.7-plus with Thinking off. It is running under a 1200-request cap with no pay-as-you-go fallback. Results and independent human adjudication are still pending; this PR is draft while retesting proceeds.